AI search is optimising the wrong layer

The next content advantage won’t come from producing more machine-readable pages. It will come from using AI to build stronger evidence, sharper narratives and a reputation worth retrieving.

I think we may be making a familiar mistake with AI search.

A new distribution system appears.

We study how it works.

Then we begin optimising everything around it.

For traditional search, that gave us an enormous SEO industry. Much of it was valuable. Search forced organisations to think about structure, discoverability, information architecture and what people were actually looking for.

But it also produced a less useful industry around the edges.

Articles created because a keyword existed.

Near-identical explainers competing for the same query.

Pages stretched to ridiculous lengths because somebody believed Google rewarded them.

Content designed primarily to satisfy an inferred machine preference rather than because an organisation had anything particularly useful to say.

AI search now risks creating its own version of that behaviour.

The vocabulary has changed. GEO. AEO. Entity optimisation. Citation analysis. LLM tracking. Machine-readable content. Schema. Prompt visibility.

Again, much of this is useful. If AI systems are becoming an important way people discover, compare and understand organisations, businesses clearly need to understand how they appear within them.

But I think the interesting question sits further upstream.

What if we’re getting better at optimising the content layer at exactly the moment that layer is becoming less scarce?

The wrong lesson from SEO

The obvious response to AI search is to ask:

How do we get cited?

How should we structure the page?

What entities should we reinforce?

Which formats work best?

How do we appear more frequently in AI-generated answers?

Those are legitimate questions.

But they contain an assumption that deserves more scrutiny: that the organisation already has something worth discovering.

That isn’t always true.

A technically perfect page cannot compensate for an organisation saying roughly the same thing as everybody else.

Schema cannot create expertise.

An FAQ cannot manufacture customer evidence.

An entity strategy cannot give a business an original point of view.

And making an unremarkable idea easier for a machine to retrieve doesn’t make the idea more remarkable.

This is where I think AI search needs to be considered as an editorial problem as well as a technical one.

I work from a fairly simple hierarchy in Editorial Intelligence:

Knowledge

↓

Narrative

↓

Communication

↓

Content

Content is the visible layer.

It is also increasingly the cheapest layer.

Generative AI can already produce competent articles, summaries, explainers, landing-page copy, FAQs and social posts at extraordinary speed. It can also restructure them, rewrite them, extract entities and create dozens of variations.

So I don’t think the long-term competitive advantage is going to come from becoming exceptionally good at producing more technically competent content.

Everybody gets those tools.

The bottleneck moves.

AI may punish the kind of content AI makes easiest to produce

There is a strange tension at the centre of the current content conversation.

AI has made generic information dramatically cheaper to produce.

AI search simultaneously reduces the need for people to visit multiple sources containing that generic information.

That should make us rethink what we’re producing.

Imagine twenty organisations publishing competent versions of “10 ways AI can improve accounts payable.”

An answer engine doesn’t necessarily need twenty versions of that article to help somebody understand the topic. It can synthesise the underlying information.

Creating article number twenty-one, even if it is beautifully structured for AI retrieval, may not give the organisation much additional strategic value.

Now imagine something different.

A company analyses two years of customer conversations and discovers that finance teams aren’t primarily worried about automation replacing jobs. Their bigger concern is losing the informal checks that currently prevent mistakes.

It investigates the pattern. It talks to customers. It compares it with product data. It commissions research. It develops a defensible position around what it calls the verification gap.

Customers begin using the phrase.

Executives discuss it.

Research supports it.

Journalists quote it.

Customer stories demonstrate it.

Sales teams hear it in conversations.

Other people begin referring to the idea.

Now there is something for an AI system to discover.

Not simply a page.

A position.

That is a very different kind of optimisation.

We need harder-edge content

This is why I suspect AI search should push content towards a harder edge.

Less generic explanation. More:

proprietary evidence
customer experience
original research
practitioner knowledge
distinctive frameworks
strong arguments
specific expertise
language captured from real conversations
claims that can actually be supported
ideas developed consistently enough that an organisation becomes associated with them

This doesn’t mean every company needs to invent a grand new category every month.

It means the starting point changes.

Instead of asking:

What content should we make about this keyword?

You start asking:

What do we actually know about this problem that is useful, credible or different?

And if the answer is “not much”, that is useful information too.

Perhaps the next step isn’t another article.

Perhaps the next step is to learn something.

The frontier AI firms offer an interesting clue

There is another reason I think this matters.

Look at what some of the companies building frontier AI are currently hiring humans to do.

OpenAI has advertised for an Executive Programs Narrative Lead whose role includes turning technology, customer insight and enterprise strategy into C-suite narratives. It is also hiring around enterprise customer storytelling, with responsibilities including interviewing customers, identifying high-impact stories, translating technical implementations and applying editorial judgement to complex inputs. Its Head of Communications, Business role explicitly includes shaping enterprise narratives, advising senior leadership and bringing real-world customer impact to life.

Anthropic, meanwhile, is listing communications roles spanning enterprise and platform alongside other editorial-oriented positions.

I wouldn’t overinterpret a handful of job descriptions. I’ve written separately about why these adverts look less like anecdotes the longer you watch them, and the caveats there apply here too.

But the relevant point for AI search is narrower.

The organisations closest to frontier AI don’t appear to be behaving as though the remaining human communications problem is simply making pages more optimisable for machines.

They are still investing in people who can understand complex technology, find the important customer evidence, shape a position, translate complexity, exercise judgement, build narratives, help executives communicate and decide what deserves attention.

That makes sense to me.

The value of those people isn’t that they can type sentences faster than an AI model.

Increasingly, they can’t.

Their value sits higher up the chain.

What matters?

What do we believe?

What can we prove?

What’s actually different?

What should somebody understand after encountering us?

What story connects all of this?

Those are harder questions.

But there is a problem: doing this manually is expensive

This is where I think the anti-AI version of the argument breaks down.

It would be easy to conclude that we should forget the machines and return to handcrafted human storytelling.

I don’t believe that.

In fact, I think we need to use AI much more aggressively. Just for different work.

Consider what exists inside a reasonably large organisation:

customer interviews
sales calls
support conversations
research reports
product discussions
meeting transcripts
event recordings
survey responses
analyst reports
performance data
executive conversations
internal documents
communities
social responses
years of published content

A traditional editorial approach asks humans to somehow keep track of all of this.

Perhaps somebody reads the research. Somebody else attends the customer meeting. A marketer remembers something an executive said six months ago. A salesperson hears the same phrase from three customers but never records it anywhere useful. A customer story is published and effectively disappears into the archive.

Then, three months later, the content team sits down to brainstorm ideas.

This is absurdly inefficient.

The organisation may already possess much of the material needed to produce genuinely distinctive thinking.

It simply cannot see it.

This is where AI changes the economics

AI can help organisations consider far more evidence than a human team could realistically process manually.

That is where I think its more interesting editorial application begins.

Not:

Write me 30 articles.

But:

Show me the recurring problems across these customer conversations.
Find language customers repeatedly use that doesn’t appear in our marketing.
Compare what customers are saying now with what they were saying six months ago.
Find evidence that supports this argument. Now find evidence that contradicts it.
Connect this new interview to relevant research we’ve already conducted.
Show me where we’re making claims without enough support.
Find expertise in these transcripts that has never been turned into formal knowledge.
Identify emerging signals that deserve investigation.
Retrieve the strongest customer evidence connected to this narrative.

That doesn’t remove the human. It changes where human attention goes.

A principle I’ve been developing within Editorial Intelligence is:

AI captures and surfaces. Humans interrogate and judge.

AI can perform retrieval, clustering, summarisation and first-pass synthesis across amounts of information that would be impractical to review manually.

Human attention can move towards interpretation, verification, connection, prioritisation and judgement.

That matters because businesses don’t have unlimited editorial time. Very few organisations are going to employ a team of people to spend weeks manually reading hundreds of transcripts in case an interesting pattern is hiding inside them.

AI makes that level of organisational listening economically plausible.

And that may ultimately produce much better content than asking AI to write the content itself.

From content engine to evidence engine

This leads me to a different model of AI-enabled content marketing.

For years, organisations have tried to build content engines. Feed ideas in. Publish assets out. Measure performance. Repeat.

AI makes those engines vastly more productive.

But productivity isn’t necessarily the constraint anymore. If everybody can produce more, producing more stops being particularly interesting.

I think the more useful system may be an Evidence Engine.

It starts with signals: customer conversations, research, behaviour, expertise, market developments, performance, external evidence.

AI helps capture, retrieve and connect those signals.

Humans decide which are meaningful. Weak observations are tested. Patterns are corroborated. Claims become stronger. Knowledge accumulates.

Then narrative gives that knowledge structure.

Instead of every article starting again from zero, individual pieces reinforce a larger argument.

That’s what I mean by Narrative Architecture.

A content calendar answers:

What are we publishing next?

Narrative Architecture asks:

What is everything we’re publishing building towards?

That distinction becomes more important in AI search because isolated optimisation creates isolated signals.

Connected evidence can create something closer to authority.

Perhaps the thing being optimised is reputation

This is the part of AI search I find most interesting.

We’ve inherited a page-centric mental model from SEO. We think about the page we want to rank. The article. The landing page. The URL.

But answer engines can encounter organisations through many kinds of evidence: owned content, customer stories, independent journalism, reviews, research, communities, executive commentary, public documentation, third-party databases and other people talking about the company.

So perhaps the strategic unit we’re trying to improve isn’t ultimately the individual webpage.

Perhaps it is the reputation a machine can infer from the evidence surrounding an organisation.

That changes the problem considerably.

You can’t simply optimise your own claim that you’re an expert. You need evidence of expertise.

You can’t simply describe yourself as trusted. Trust needs signals.

You can’t declare yourself the authority on a problem and expect the rest of the information environment to agree.

Reputation has to accumulate.

This is why I think about AI visibility and buyability as connected problems.

Visibility asks whether people and AI systems can find and understand you.

Buyability asks whether enough credible evidence exists for somebody to feel confident choosing you.

Those things connect. Evidence builds reputation. Reputation supports recommendation. Recommendation helps reduce the uncertainty involved in buying.

Technical visibility matters throughout that process.

But technical visibility alone doesn’t create the underlying confidence.

So yes, optimise

None of this is an argument against GEO.

I expect technical AI-search work to become a normal part of digital publishing.

Make important information accessible. Structure it clearly. Make entities understandable. Maintain good technical foundations. Give machines clean access to genuine expertise. Understand where your organisation appears and where it doesn’t. Measure it. Improve it.

But I would be wary of treating that as the whole strategy.

Because optimisation becomes much more powerful once an organisation has built something worth optimising.

The harder competitive question is upstream:

What does the organisation know?

What evidence does it possess?

What is it learning from customers?

What expertise is currently trapped inside people and conversations?

Which assumptions should be challenged?

Which ideas can it genuinely own?

What story connects its evidence?

What does it want to become known for?

AI can help answer those questions at a scale that wasn’t previously practical.

That’s the opportunity I think we’re underestimating.

Use AI higher up the value chain

The mistake isn’t using AI too much.

It’s using AI too low down the value chain.

If we use generative AI mainly to recreate the SEO content machine more cheaply, we may end up with an industrial-scale version of the least interesting part of the old internet.